Why remove the effects of metadata?

[kpmokpmo]

Hi,

Thank you for your work. But I am a little bit confused. Since race plays a role in predicting gender why not including them in the input data? I know that these data are usually from different modalities, but we can do early/late fuse or whatever to explicitly take them in. I am not sure if there will be some negative impact if data are not very orthogonal to each other in deep learning framework.

I also come up with another paper "REVERSIBLE INSTANCE NORMALIZATION FOR ACCURATE TIME-SERIES FORECASTING AGAINST DISTRIBUTION SHIFT"(under review). In that work, they get rid of the instance-wise mean of the initial time series and plug it back right before the output layer. Can I think of the instance-wise mean as meta data in this sense?

Thanks!